In this case it's not just a library like the wrappers for tensorflow already present in many languages (including python), it's something like a Swift language fork in which the language constructs are analyzed by the forked compiler during compilation in order to directly generate an equivalent tensorflow execution graph (without manually creating a tape through library calls). Thus it's a Swift (compiler plugin) for tensorflow.

The intention was to merge the fork into the main language making Swift differentiation tools first class within the language.